A method and system for full-field stress digital twin reconstruction of a spatial docking mechanism oriented to sparse sensor conditions

Abstract

The application provides a space docking mechanism full-field stress digital twin reconstruction method and system for sparse sensor conditions, belonging to the technical field of spacecraft structure health monitoring. The method: acquires fine grid node stresses of a high-precision finite element model of a space docking mechanism under each load condition and corresponding load conditions; maps the fine grid node stresses to rough nodes to obtain rough node stress data sets; dimensionally reduces and samples the rough node stress data sets to concentrate samples in high-stress intensity areas and outputs dimensionally reduced sample sets; constructs the dimensionally reduced sample sets into graph structure data sets; trains a graph neural network model to obtain a virtual sensor model; and deploys the virtual sensor model in a digital twin system, inputs real-time sparse loads, outputs full-field node stress prediction values, and realizes real-time reconstruction and monitoring. The application can realize high-precision real-time reconstruction of mechanism full-field stresses under sparse sensor conditions.

Description

Technical Field

[0001] This application belongs to the field of spacecraft structural health monitoring technology, specifically relating to a digital twin reconstruction method and system for full-field stress of space docking mechanisms under sparse sensor conditions. Background Technology

[0002] SDM is short for Space Docking Mechanism.

[0003] The SDM is the core component for achieving on-orbit docking and separation of spacecraft. Its structural reliability directly determines the safety of rendezvous and docking missions and is a key technological guarantee for complex space operations such as building large space stations and conducting deep space exploration.

[0004] However, constrained by factors such as on-orbit microgravity, high vacuum, and extreme alternating thermal environment, the SDM structure is extremely compact, making it difficult to densely deploy physical sensors in key stress concentration areas. The actual monitoring data obtainable comes only from sparse load information at a few measuring points, failing to cover most of the structure. This inherent constraint poses a significant challenge to directly acquiring the real-time mechanical response of the SDM across the entire field. Existing engineering practices mainly rely on pre-launch ground tests or offline analysis based on limited measuring points, which cannot comprehensively characterize the mechanism's actual on-orbit evolution state, nor can they meet the urgent need for real-time monitoring of structural integrity and reconstruction of the entire field state during docking.

[0005] To address the aforementioned issues, digital twins have been introduced for synchronous mapping between virtual and real systems. Existing SDM mechanical performance evaluation methods can be categorized into three types: The first is data-driven methods, which treat data as regular grids or time series, ignoring the inherent spatial topology of the SDM multibody system and making it difficult to characterize the irregular spatial relationships between nodes; the second is mechanism-driven methods (such as finite element analysis), which, while physically interpretable, are complex to model and computationally expensive, failing to meet the millisecond-level real-time response requirements of digital twins; the third is hybrid data and mechanism methods, which, although proven effective, still lack a unified technical solution for fully integrating the SDM node topology relationships with small sample data to construct a high-fidelity real-time stress reconstruction model under sparse sensor conditions.

[0006] Therefore, there is an urgent need for a digital twin reconstruction method for the full-field stress of space docking mechanisms under sparse sensor conditions to overcome the shortcomings of the existing technologies. Summary of the Invention

[0007] In a first aspect, embodiments of this application provide a method for full-field stress digital twin reconstruction of a space docking mechanism under sparse sensor conditions, comprising the following steps: S1. Obtain the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition; S2. Map the stress at the fine mesh nodes to the coarsened nodes to obtain the coarsened node stress dataset; S3. Perform dimensionality reduction sampling on the coarsened nodal stress dataset to concentrate the samples in the region of high stress intensity, and output the dimensionality-reduced sample set; S4. Construct the dimensionality-reduced sample set into a graph structure dataset. The node features in the graph structure dataset include load conditions and spatial coordinates. The node labels are stress values. The edges between nodes are determined based on the finite element mesh connection relationships. S5. Construct a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model; S6. Deploy the virtual sensor model in the digital twin system, take the sparse loads collected in real time by the space docking mechanism as input, and output the predicted values ​​of stress at all nodes in the field to realize real-time reconstruction and monitoring of stress in the field.

[0008] Secondly, embodiments of this application also provide a full-field stress digital twin reconstruction system for space docking mechanisms under sparse sensor conditions, comprising: The data acquisition module is used to acquire the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition. The stress mapping module is used to map the stress of fine mesh nodes to coarse nodes, resulting in a coarsened node stress dataset. The dimension reduction sampling module is used to perform dimension reduction sampling on the coarsened nodal stress dataset, so that the samples are concentrated in the region of high stress intensity, and output the dimension-reduced sample set. The graph structure dataset construction module is used to construct a graph structure dataset from the dimensionality-reduced sample set. The node features in the graph structure dataset include load conditions and spatial coordinates, the node label is the stress value, and the edges between nodes are determined according to the finite element mesh connection relationship. The model training module is used to build a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model. The real-time reconstruction and monitoring module is used to deploy virtual sensor models in the digital twin system. It takes the sparse loads collected in real time by the space docking mechanism as input and outputs the predicted values ​​of stress at all nodes in the field, so as to realize the real-time reconstruction and monitoring of stress in the field.

[0009] As can be seen from the above technical solutions, this application has the following advantages: The method and system for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions provided in this application resolves the contradiction between the large amount of node data in high-precision finite element models and the real-time response requirements of digital twins by mapping the nodal stress in high-precision finite element models to coarsened nodes and performing dimensionality reduction sampling, combined with the ability of graph neural network models to represent spatial topological relationships. This application overcomes the inherent defects of space docking mechanisms, which are difficult to densely arrange physical sensors due to their compact structure and limited space, and cannot directly obtain the real-time stress response of the entire field. Under the condition of only obtaining the triaxial force and triaxial moment as sparse loads, high-precision real-time reconstruction of full-field stress can be achieved, providing integrity monitoring for on-orbit docking structures. Attached Figure Description

[0010] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating the full-field stress digital twin reconstruction method for space docking mechanisms under sparse sensor conditions according to the present invention.

[0012] Figure 2 This is a schematic diagram of the finite element nodal stress mapping process in this application.

[0013] Figure 3 This is a schematic diagram of the mechanical performance prediction process for the space docking mechanism based on GCN-DNN in this application.

[0014] Figure 4 This is a schematic diagram of the workflow of the digital twin real-time stress reconstruction system of this application.

[0015] Figure 5 This is a schematic diagram of the interface of the digital twin mechanical performance prediction system for the space docking mechanism in this application.

[0016] Figure 6 This is a schematic diagram of the full-field stress digital twin reconstruction system for space docking mechanisms under sparse sensor conditions according to the present invention. Detailed Implementation

[0017] Various embodiments of this disclosure will be described more fully in the following detailed description of the specific steps of the full-field stress digital twin reconstruction method for space docking mechanisms under sparse sensor conditions. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0018] This embodiment provides a digital twin reconstruction method for the full-field stress of a space docking mechanism under sparse sensor conditions, which can achieve high-precision real-time reconstruction and monitoring of the full-field stress of the docking mechanism under the condition of only obtaining sparse loads.

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 The diagram shows a flowchart of a digital twin reconstruction method for the full-field stress of a space docking mechanism under sparse sensor conditions, according to a specific embodiment. The method includes the following steps: S1. Obtain the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition; It should be noted that by obtaining the nodal stress under fine mesh, the resolution and accuracy of the training data are ensured, providing a benchmark reference and data source for high-precision reconstruction under sparse sensor conditions; S2. Map the stress at the fine mesh nodes to the coarsened nodes to obtain the coarsened node stress dataset; It should be noted that by mapping, the coarsened dataset is obtained, which reduces the data dimensionality while retaining the main mechanical features, enabling real-time calculation and balancing accuracy and efficiency. S3. Perform dimensionality reduction sampling on the coarsened nodal stress dataset to concentrate the samples in the region of high stress intensity, and output the dimensionality-reduced sample set; It should be noted that this step uses stress intensity as the sampling target, concentrating the samples in key areas where stress is concentrated and failure is likely to occur. This retains key mechanical information while compressing the size of the dataset. There is still a lot of redundant information between the coarsened nodes. If all of it is used for model training, it will introduce a lot of redundant computation. Through this step of dimensionality reduction sampling, the number of training samples is reduced without significantly affecting the model accuracy, thus improving the efficiency of model training. S4. Construct the dimensionality-reduced sample set into a graph structure dataset. The node features in the graph structure dataset include load conditions and spatial coordinates. The node labels are stress values. The edges between nodes are determined based on the finite element mesh connection relationships. It should be noted that this step, by explicitly constructing a graph structure, enables the model to aggregate features using the topological connections between nodes, thereby more accurately representing the stress transmission path inside the mechanism. S5. Construct a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model; It should be noted that this step utilizes the inherent characteristics of graph neural networks in processing graph structure data. It can capture the topological relationships between nodes in non-Euclidean space by aggregating information from neighboring nodes. Combined with deep neural networks, it can establish a high-precision mapping from sparse loads to full-field stress. After training, the model can output the predicted values ​​of full-field nodal stress in real time with only a six-dimensional sparse load as input, without relying on finite element calculations or dense sensor arrays, thus reducing the data requirements and computational latency in the online monitoring stage. S6. Deploy the virtual sensor model in the digital twin system, take the sparse loads collected in real time by the space docking mechanism as input, and output the predicted values ​​of stress at all nodes in the field to realize real-time reconstruction and monitoring of stress in the field. It should be noted that this step deploys the virtual sensor model on the digital twin system. By taking the sparse loads collected in real time by the physical docking mechanism as input, the predicted values ​​of the stress at all nodes in the field can be output in real time and visualized. This realizes a closed loop from offline simulation modeling to online real-time monitoring, and solves the problems of large calculation delay and inability to respond in real time in the traditional finite element method. It provides real-time decision support information for structural integrity monitoring and anomaly early warning during the docking process.

[0021] This embodiment overcomes the limitations of sensor deployment and measurement point arrangement caused by the compact structure by mapping fine mesh stress and using graph neural networks to train virtual sensor models. It can achieve millisecond-level real-time reconstruction and visualization monitoring of full-field stress under the condition of only acquiring sparse loads, thus ensuring on-orbit docking safety.

[0022] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process in this embodiment, another method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions is provided. This method includes the following steps: S1. Obtain the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition; The specific steps of step S1 are as follows: S11. Establish a high-precision finite element model of the space docking mechanism; The high-precision finite element model is established based on the actual geometric dimensions of the space docking mechanism. The mesh density meets the requirements for capturing the stress gradient in key areas. Typically, the fine mesh node scale reaches millions to ensure that the accuracy of the simulation results meets the requirements of subsequent digital twin modeling. S12. Apply different load conditions, the load conditions including triaxial forces Fx, Fy, Fz and triaxial moments Mx, My, Mz; The load conditions cover various typical load combinations that may occur during the docking process; the three forces Fx, Fy, and Fz represent the forces acting along the three orthogonal directions in space, and the three moments Mx, My, and Mz represent the moments about the three orthogonal directions; by applying the above six-dimensional load boundary conditions to the space docking mechanism, it is possible to comprehensively simulate the various mechanical loads it bears during the on-orbit docking process. S13. Solve the high-precision finite element model and obtain the stress response Von Mises at each node of the fine mesh; The finite element model is solved to obtain the Von Mises equivalent stress response of each fine mesh node. Von Mises stress is an equivalent stress index that comprehensively considers multiple stress components. It can effectively characterize the yield tendency of materials under complex stress states and is a key mechanical parameter widely used in structural strength assessment. S14. Associate and store the load conditions corresponding to each load case with the stress of the corresponding fine mesh nodes to form the original data source; The six-dimensional load conditions corresponding to each load case are associated and stored with the corresponding fine mesh node Von Mises stress, forming a structured raw data source, which provides a standardized data input format for subsequent steps; S2. Map the stress at the fine mesh nodes to the coarsened nodes to obtain the coarsened node stress dataset; The specific steps of step S2 are as follows: S21. Use fine mesh nodes as the source node set and coarse nodes of a preset size as the target node set; The preset size of the coarsening nodes needs to strike a balance between stress field fidelity and computational efficiency. The number of target nodes should be sufficient to preserve the key geometric features and stress distribution patterns of the structure, while also meeting the node size constraints of real-time rendering of digital twins. S22. Construct a spatial index and determine the neighboring nodes and candidate tetrahedral elements of each coarsened node in the fine mesh through nearest neighbor search; In step S22, a Kd-tree spatial index is constructed using the coordinates of all fine-grained grid nodes, and the space is recursively divided by binary search according to the dimension of maximum variance. Kd-tree is an efficient binary tree spatial index structure that can perform nearest neighbor search in high-dimensional space with an average time complexity of O(log N), which is far superior to O(N) linear search that traverses all nodes. However, due to the large scale of fine-grid nodes, without an efficient search strategy, locating each coarsened node will consume a lot of computational resources, which will seriously affect the overall efficiency. For each coarsened node, the nearest fine-grained node is located using a nearest neighbor search with minimum Euclidean distance. The nearest neighbor search formula is:

[0023] in, Represents a set of fine-grained grid nodes The i-th node in Represents fine mesh nodes spatial coordinates, Represents the spatial coordinates of the target coarsening node. This represents the finest node that is closest to the target coarsening node; S23. Determine the positional relationship between the coarsened node and the candidate tetrahedral element based on the centroid coordinate method, and locate the tetrahedral element to which the coarsened node belongs; In step S23, let the four vertices of the candidate tetrahedral element be V1, V2, V3, and V4. The centroid weight of the coarsened node relative to the candidate tetrahedral element is calculated by solving the following system of linear equations. :

[0024] in, These are the spatial coordinates of the four vertices of the tetrahedron. Coarse out the spatial coordinates of the target node; Based on the weight of the center of gravity Determine the positional relationship by the sign: when and When the coarsened node is located inside the tetrahedron, it is determined that the node is located inside the tetrahedron. When it exists and When the roughened node is located on the tetrahedral surface, it is determined that the roughened node is located on the tetrahedral surface. When it exists and If i, j, k, and l are all different, then the coarsened node is located on the edge of the tetrahedron. When it exists And the remaining When both are 0, the coarsened node is determined to be located at a tetrahedron vertex; When it exists or When the coarsened node is located outside the tetrahedron, it is determined that the node is located outside the tetrahedron. The barycentric coordinate method is a classic element inclusion test method in computational geometry; The geometric meaning of is the volume coordinate of the coarsened node relative to the i-th vertex of the tetrahedron, and its value reflects the relative position of the coarsened node in the directions of each vertex of the tetrahedron; by analyzing... Determining the sign and value range can accurately determine the spatial positional relationship between the coarsened node and the tetrahedral element, which is the basis for displacement interpolation based on shape function. If the coarsening node is outside the current tetrahedron, then iteratively check the adjacent tetrahedrons formed by the faces of the current tetrahedron; If, according to the above centroid coordinate determination rules, the coarsened node is located inside, on the surface, on the edge, or at the vertex of a certain adjacent tetrahedron, then this adjacent tetrahedron is taken as the located tetrahedral element, and the process proceeds to step S24 for displacement interpolation. If the target coarsening node is not located after traversing all adjacent tetrahedrons, then the nearest neighbor node in step S22 is directly selected. The stress value is assigned to the target coarsening node; The above-mentioned adjacent tetrahedron traversal strategy ensures the robustness of the algorithm. For nodes that cannot be accurately located due to irregular mesh or coarsened nodes located at cell boundaries, traversing adjacent cells can effectively solve the problem of cross-cell boundary positioning. If the node is still not located after traversal due to numerical error or mesh distortion, the stress value of the nearest neighbor node is used as an alternative to prevent the entire mapping process from being interrupted due to the failure of individual node positioning. S24. Use the shape function of the tetrahedral element to perform displacement interpolation to obtain the displacement of the coarsened node; In step S24, the shape function of the linear tetrahedral element is used. Perform displacement interpolation, the shape function Numerically equal to the centroid weight obtained in step S23. Displacement of coarsened nodes Calculate using the following formula:

[0025] Where m represents the vertex index. This represents the displacement vector of the m-th vertex of the tetrahedron; For linear tetrahedral elements, their shape functions are numerically identical to their volume coordinates (i.e., barycentric weights), therefore the values ​​obtained in step S23 can be directly used. Interpolation can be performed without calculating shape functions separately, simplifying the calculation process; this interpolation method ensures the continuity of the coarsened nodal displacement field, which is consistent with the finite element method. S25. Based on the material constitutive equation, the stress values ​​of the coarsened nodes are inversely derived from their displacements to obtain the coarsened node stress dataset. In step S25, the material constitutive equation adopts Hooke's law to coarseen the stress at the nodes. Calculate using the following formula:

[0026] in, This represents the stress value of the coarsened node. Represents the elasticity matrix of the material. Let be the displacement gradient tensor matrix. The coarsened nodal displacements obtained through interpolation; By using Hooke's law to back-calculate the stress, the mapped stress field strictly satisfies the constitutive relation of continuous medium mechanics, thus ensuring the physical consistency of the stress mapping. like Figure 2 As shown, the process constructs a Kd-tree spatial index using all fine-grid nodes and performs a nearest-neighbor search to determine the neighboring nodes and candidate tetrahedral elements of each coarsened node in the fine-grid. Then, based on the centroid coordinate method, the positional relationship between the coarsened node and the candidate tetrahedral elements is determined, locating the tetrahedral element to which the target coarsened node belongs. Next, displacement interpolation is performed using the shape function of the linear tetrahedral element, obtaining the displacement of the coarsened node from the four vertex displacements. Finally, based on Hooke's law constitutive equation, the stress value of the coarsened node is calculated using the displacement gradient tensor and the elasticity matrix, completing the mapping from the fine-grid to the high-precision stress field to the stress dataset of the coarsened node. If the coarsened node is located outside the current tetrahedron, the adjacent tetrahedrons are traversed iteratively. If it is still not located after traversal, the stress value of the nearest neighbor node is directly assigned to the coarsened node to ensure the robustness of the algorithm. S3. Perform dimensionality reduction sampling on the coarsened nodal stress dataset to concentrate the samples in the region of high stress intensity, and output the dimensionality-reduced sample set; The specific steps of step S3 are as follows: S31. The probability density function of the coarsened nodal stress dataset is fitted using kernel density estimation to obtain the target probability distribution; In step S31, a Gaussian kernel function is used to estimate the kernel density, and the estimated probability density of any stress value s is obtained. Calculate as follows:

[0027] Among them, Gaussian kernel function for:

[0028] In the formula, This represents the total number of nodes after coarsening. This represents the normalized stress value at the i-th node. The bandwidth parameter is defined as the standard deviation of the stress data; the target distribution is defined as... ; Kernel density estimation, as a non-parametric estimation method, does not require prior assumption that stress data follows a specific distribution (such as a normal distribution). It can adaptively fit probability density functions of arbitrary shapes, ensuring the accuracy of the target distribution. The bandwidth parameter h is taken as the standard deviation of the stress data. Its value can balance the smoothness and detail fidelity of probability density estimation. If it is too large, it will be over-smoothed and lose key information. If it is too small, it will introduce unnecessary noise. S32. Using stress intensity as the sampling target, construct a Markov chain and use the Metropolis-Hastings algorithm to sample on the target distribution, so that the sampling nodes are concentrated in the region of high stress intensity. In step S32, the acceptance probability of the Metropolis-Hastings algorithm Calculate using the following formula:

[0029] in, Indicates the current state. Indicates the candidate state. Indicates the target distribution. Indicates the suggested distribution; Generate random numbers u from a uniform distribution U(0,1); like Then accept the candidate state. As a new state ; like If the candidate state is not found, the state remains unchanged. ; Markov chains satisfy the following detailed equilibrium conditions:

[0030] in, Indicates starting from the current state Transition to candidate state The probability of; The acceptance probability calculation formula of the Metropolis-Hastings algorithm ensures that the Markov chain satisfies the detailed equilibrium condition, which theoretically guarantees that the stable distribution of the chain is exactly equal to the target distribution, thus providing the sampling results with strict mathematical theoretical protection. By associating the sampling probability with the stress intensity, high-stress areas are densely sampled and low-stress areas are sparsely sampled, thereby significantly compressing the size of the dataset while preserving the key mechanical topological features of the docking mechanism. S33. Introduce a local perception candidate generation strategy based on spatial indexing to replace the global uniform candidate, and generate candidate states from the spatial neighborhood of the current state; A local sensing candidate generation strategy is constructed based on Kd-tree spatial indexing. Candidate states are generated from the spatial neighborhood of the current state in the Kd-tree, effectively concentrating sampling in high-stress gradient regions and improving the convergence speed and sampling efficiency of Markov chains. The standard Metropolis-Hastings algorithm uses globally uniform candidates, which may appear in low-probability regions far from the current state, resulting in a large number of proposals being rejected and low sampling efficiency. This step restricts candidate generation to the spatial neighborhood of the current state through Kd-trees, making the Markov chain walk concentrated in high-stress gradient regions, improving the acceptance rate and convergence speed, and enhancing the overall efficiency of dimensionality reduction sampling. S34. Once the Markov chain reaches a stationary distribution, output the dimensionality-reduced sample set. S4. Construct the dimensionality-reduced sample set into a graph structure dataset. The node features in the graph structure dataset include load conditions and spatial coordinates. The node labels are stress values. The edges between nodes are determined based on the finite element mesh connection relationships. The specific steps of step S4 are as follows: S41. Encode the load conditions and spatial coordinates of each sampling node in the dimensionality-reduced sample set into a feature vector. The feature vector is a 9-dimensional vector, including triaxial forces Fx, Fy, Fz, triaxial moments Mx, My, Mz, and spatial coordinates x, y, z. The 9-dimensional feature vector encodes the load conditions and spatial coordinates in a unified manner, enabling the model to simultaneously perceive external load information and node spatial location information, providing a complete input feature space for graph neural networks; S42. Use the stress response Von Mises of the sampling node under the corresponding load condition as the label of the sampling node; S43. Define the edges between sampling nodes based on the actual connection relationships of the finite element mesh. Nodes belonging to the same finite element element are connected to form a graph structure. ; in, Represents a set of nodes. Represents the set of edges; The adjacency matrix of the graph structure G Defined as:

[0031] The adjacency matrix is ​​defined based on the co-ownership relationship of the finite element mesh, ensuring that each edge in the graph structure corresponds to a real stress transmission path, rather than an artificial connection based on a distance threshold, thus fundamentally guaranteeing the physical authenticity of the graph topology. This construction method enables the subsequent graph convolutional network to aggregate the stress response characteristics between mechanically relevant nodes when aggregating neighbor node information. S44. Divide the graph structure dataset into training, validation and test sets proportionally; The training set is used for learning and updating model parameters, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used to evaluate the model's generalization performance on unseen data. The datasets divided according to the above proportions provide a standardized data partitioning scheme for model training, validation, and testing. S5. Construct a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model; In step S5, the graph neural network model is a hybrid model of a cascaded graph convolutional network and a deep neural network; the specific steps are as follows: S51. Normalize the node features in the graph structure dataset to eliminate dimensional differences; specifically, Z-score standardization is applied to spatial coordinates, and coordinate sequences are normalized. Calculate the standardized value using the following formula. :

[0032] in, Represents the mean of the sample sequence. Indicates the standard deviation of the sample sequence; Z-score standardization can eliminate the influence of differences in units and orders of magnitude between different features, prevent features with larger values ​​from dominating model training, and preserve the physical meaning of negative coordinate values. It is suitable for standardization of spatial coordinates. S52. Input the normalized node features and graph structure into the graph convolutional network. The graph convolutional network extracts local spatial topological features by aggregating neighbor node information through topology, and obtains the graph structure feature representation. Graph convolutional networks can perform convolution operations on graph-structured data, extracting the hidden features of nodes in the local topology by aggregating the feature information of each node and its neighboring nodes. Graph convolutional networks adopt a two-layer structure, which can aggregate second-order neighbor information and capture a sufficient range of local topological features while controlling the complexity of the model. S53. Perform an embedding average pooling operation on the graph structure feature representation to compress the features of nodes and their neighbors into dense latent vectors, which are then input into a deep neural network; the embedding average pooling is calculated according to the following formula:

[0033] in, Let k represent the feature representation of a node and its neighbors after graph convolution, where k represents the number of neighboring nodes and h represents the dimension of the latent vector. This represents the dense latent vector obtained after compression; Embedded average pooling transforms the irregular output of graph structures into fixed-length vector representations, realizing the connection between graph neural networks and deep neural networks, enabling subsequent fully connected layers to process the graph structure features output by graph convolutional networks; S54. The deep neural network performs regression prediction on the pooled dense latent vectors and outputs the predicted values ​​of node stress. The deep neural network has a four-hidden-layer funnel structure with dimensions of 256, 128, 64, and 32 respectively. Each hidden layer uses the ReLU activation function. The four-hidden-layer funnel structure forces the network to learn a hierarchical feature abstraction from a high-dimensional latent space to a low-dimensional stress output by reducing dimensionality layer by layer. The use of the ReLU activation function in each hidden layer can effectively alleviate the gradient vanishing problem and accelerate convergence. like Figure 3 As shown, the process first performs Z-score normalization on the 9-dimensional node features in the graph structure dataset to eliminate dimensional differences. Then, the normalized node features and the graph structure are input into a graph convolutional network (PCN). The PCN extracts local spatial topological features by aggregating neighbor node information through topology aggregation. Subsequently, the output of the PCN is subjected to embedding average pooling to convert the irregular output of the graph structure into a fixed-length dense latent vector. Finally, the dense latent vector is input into a four-hidden-layer funnel-shaped deep neural network for regression prediction, outputting the stress prediction value of each node. The entire model uses mean squared error (MSE) as the loss function and is trained using end-to-end backpropagation, enabling joint optimization of the graph convolutional network and the deep neural network. S6. Deploy the virtual sensor model in the digital twin system, take the sparse load collected in real time by the space docking mechanism as input, and output the predicted value of the stress at all nodes in the field to realize the real-time reconstruction and monitoring of the stress in the field. The specific steps of step S6 are as follows: S61. Deploy the trained virtual sensor model into the digital twin system; S62. The space docking mechanism synchronizes with the digital twin model through real-time data interaction, and inputs the triaxial forces Fx, Fy, Fz and triaxial torques Mx, My, Mz collected in real time as sparse loads into the virtual sensor model; S63. The virtual sensor model reconstructs the stress of each node in the entire field based on the real-time sparse load, and visualizes the reconstructed stress of each node in the three-dimensional scene in real time, so as to realize the real-time monitoring and anomaly warning of the structural integrity during the docking process, and provide assistance for on-orbit docking safety decision-making. like Figure 4As shown, after the trained virtual sensor model is deployed in the digital twin system, it enters the online operation phase. The spatial docking mechanism in the physical space keeps synchronized with the digital twin model through real-time data interaction, inputting the real-time collected sparse loads, namely the triaxial forces Fx, Fy, Fz and the triaxial moments Mx, My, Mz, into the deployed virtual sensor model. After receiving the sparse loads, the virtual sensor model completes inference within milliseconds and outputs the stress prediction values ​​of each node in the entire field. The reconstructed full-field stress data is visualized and rendered in real-time in the form of cloud maps in the three-dimensional scene, realizing the real-time monitoring and anomaly warning functions of the structural integrity during the docking process.

[0034] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0035] like Figure 6 As shown, the following is an embodiment of the full-field stress digital twin reconstruction system for space docking mechanisms under sparse sensor conditions provided by this disclosure. This system and the full-field stress digital twin reconstruction method for space docking mechanisms under sparse sensor conditions described in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the full-field stress digital twin reconstruction system for space docking mechanisms under sparse sensor conditions, please refer to the embodiments of the full-field stress digital twin reconstruction method for space docking mechanisms under sparse sensor conditions described above.

[0036] The system includes: The data acquisition module is used to acquire the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition. The stress mapping module is used to map the stress of fine mesh nodes to coarse nodes, resulting in a coarsened node stress dataset. The dimension reduction sampling module is used to perform dimension reduction sampling on the coarsened nodal stress dataset, so that the samples are concentrated in the region of high stress intensity, and output the dimension-reduced sample set. The graph structure dataset construction module is used to construct a graph structure dataset from the dimensionality-reduced sample set. The node features in the graph structure dataset include load conditions and spatial coordinates, the node label is the stress value, and the edges between nodes are determined according to the finite element mesh connection relationship. The model training module is used to build a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model. The real-time reconstruction and monitoring module is used to deploy virtual sensor models in a digital twin system. It takes the sparse loads collected in real time by the physical docking mechanism as input and outputs the predicted values ​​of stress at all nodes in the field, thereby realizing real-time reconstruction and monitoring of stress in the field.

[0037] like Figure 5 As shown, the interface of the space docking mechanism digital twin mechanical performance prediction system is used to display the stress field distribution reconstructed by the digital twin system and key monitoring data during the docking process, providing operators with an intuitive display of the mechanical state.

[0038] This embodiment overcomes the inherent constraint of the docking mechanism's difficulty in densely arranging sensors by leveraging the interactive collaboration of the data acquisition module, stress mapping module, dimensionality reduction sampling module, graph structure dataset construction module, model training module, and real-time reconstruction and monitoring module. Under the condition of only acquiring sparse loads, it achieves high-precision real-time reconstruction and monitoring of stress across the entire field, meeting the real-time requirements of on-orbit digital twins.

[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions, characterized in that, Includes the following steps: S1. Obtain the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition; S2. Map the stress at the fine mesh nodes to the coarsened nodes to obtain the coarsened node stress dataset; S3. Perform dimensionality reduction sampling on the coarsened nodal stress dataset to concentrate the samples in the region of high stress intensity, and output the dimensionality-reduced sample set; S4. Construct the dimensionality-reduced sample set into a graph structure dataset. The node features in the graph structure dataset include load conditions and spatial coordinates. The node labels are stress values. The edges between nodes are determined based on the finite element mesh connection relationships. S5. Construct a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model; S6. Deploy the virtual sensor model in the digital twin system, take the sparse loads collected in real time by the space docking mechanism as input, and output the predicted values ​​of stress at all nodes in the field to realize real-time reconstruction and monitoring of stress in the field.

2. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11. Establish a high-precision finite element model of the space docking mechanism; S12. Apply different load conditions, the load conditions including triaxial forces Fx, Fy, Fz and triaxial moments Mx, My, Mz; S13. Solve the high-precision finite element model and obtain the stress response Von Mises at each node of the fine mesh; S14. Associate and store the load conditions corresponding to each load case with the stress of the corresponding fine mesh nodes to form the original data source.

3. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Use fine mesh nodes as the source node set and coarse nodes of a preset size as the target node set; S22. Construct a spatial index and determine the neighboring nodes and candidate tetrahedral elements of each coarsened node in the fine mesh through nearest neighbor search; S23. Determine the positional relationship between the coarsened node and the candidate tetrahedral element based on the centroid coordinate method, and locate the tetrahedral element to which the coarsened node belongs; S24. Use the shape function of the tetrahedral element to perform displacement interpolation to obtain the displacement of the coarsened node; S25. Based on the material constitutive equation, the stress values ​​of the coarsened nodes are inversely derived from their displacements to obtain the coarsened node stress dataset.

4. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 3, characterized in that, In step S22, a Kd-tree spatial index is constructed using the coordinates of all fine-grained grid nodes, and the space is recursively divided by binary search according to the dimension of maximum variance. For each coarsened node, the nearest fine-grained node is located using a nearest neighbor search with minimum Euclidean distance. The nearest neighbor search formula is: in, Represents a set of fine-grained grid nodes The i-th node in Represents fine mesh nodes spatial coordinates, Represents the spatial coordinates of the target coarsening node. This represents the finest node that is closest to the target coarsening node; In step S23, let the four vertices of the candidate tetrahedral element be V1, V2, V3, and V4. The centroid weight of the coarsened node relative to the candidate tetrahedral element is calculated by solving the following system of linear equations. : in, These are the spatial coordinates of the four vertices of the tetrahedron. Coarse out the spatial coordinates of the target node; Based on the weight of the center of gravity Determine the positional relationship by the sign: when and When the coarsened node is located inside the tetrahedron, it is determined that the node is located inside the tetrahedron. When it exists and When the roughened node is located on the tetrahedral surface, it is determined that the roughened node is located on the tetrahedral surface. When it exists and If i, j, k, and l are all different, then the coarsened node is located on the edge of the tetrahedron. When it exists And the remaining When both are 0, the coarsened node is determined to be located at a tetrahedron vertex; When it exists or When the coarsened node is located outside the tetrahedron, it is determined that the node is located outside the tetrahedron. If the coarsening node is outside the current tetrahedron, then iteratively check the adjacent tetrahedrons formed by the faces of the current tetrahedron; If, according to the above centroid coordinate determination rules, the coarsened node is located inside, on the surface, on the edge, or at the vertex of a certain adjacent tetrahedron, then this adjacent tetrahedron is taken as the located tetrahedral element, and the process proceeds to step S24 for displacement interpolation. If the target coarsening node is not located after traversing all adjacent tetrahedrons, then the nearest neighbor node in step S22 is directly selected. The stress value is assigned to the target coarsening node.

5. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 3, characterized in that, In step S24, the shape function of the linear tetrahedral element is used. Perform displacement interpolation, the shape function Numerically equal to the centroid weight obtained in step S23. Displacement of coarsened nodes Calculate using the following formula: Where m represents the vertex index. This represents the displacement vector of the m-th vertex of the tetrahedron; In step S25, the material constitutive equation adopts Hooke's law to coarseen the stress at the nodes. Calculate using the following formula: in, This represents the stress value of the coarsened node. Represents the elasticity matrix of the material. Let be the displacement gradient tensor matrix. This is the coarsened nodal displacement obtained through interpolation.

6. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 1, characterized in that, The specific steps of step S3 are as follows: S31. The probability density function of the coarsened nodal stress dataset is fitted using kernel density estimation to obtain the target probability distribution; In step S31, a Gaussian kernel function is used to estimate the kernel density, and the estimated probability density of any stress value s is obtained. Calculate as follows: Among them, Gaussian kernel function for: In the formula, This represents the total number of nodes after coarsening. This represents the normalized stress value at the i-th node. For bandwidth parameters; the target distribution is defined as follows: ; S32. Using stress intensity as the sampling target, construct a Markov chain and use the Metropolis-Hastings algorithm to sample on the target distribution, so that the sampling nodes are concentrated in the region of high stress intensity. In step S32, the acceptance probability of the Metropolis-Hastings algorithm Calculate using the following formula: in, Indicates the current state. Indicates the candidate state. Indicates the target distribution. Indicates the suggested distribution; Generate random numbers u from a uniform distribution U(0,1); like Then accept the candidate state. As a new state ; like If the candidate state is not found, the state remains unchanged. ; Markov chains satisfy the following detailed equilibrium conditions: in, Indicates starting from the current state Transition to candidate state The probability of; S33. Introduce a local perception candidate generation strategy based on spatial indexing to replace the global uniform candidate, and generate candidate states from the spatial neighborhood of the current state; A local sensing candidate generation strategy is constructed based on Kd-tree spatial indexing. Candidate states are generated from the spatial neighborhood of the current state in the Kd-tree, effectively concentrating sampling in high-stress gradient regions and improving the convergence speed and sampling efficiency of Markov chains. S34. Once the Markov chain reaches a stationary distribution, output the dimensionality-reduced sample set.

7. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Encode the load conditions and spatial coordinates of each sampling node in the dimensionality-reduced sample set into a feature vector. The feature vector is a 9-dimensional vector, including triaxial forces Fx, Fy, Fz, triaxial moments Mx, My, Mz, and spatial coordinates x, y, z. S42. Use the stress response Von Mises of the sampling node under the corresponding load condition as the label of the sampling node; S43. Define the edges between sampling nodes based on the actual connection relationships of the finite element mesh. Nodes belonging to the same finite element element are connected to form a graph structure. ; in, Represents a set of nodes. Represents the set of edges; The adjacency matrix of the graph structure G Defined as: S44. Divide the graph structure dataset into training, validation and test sets proportionally.

8. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 1, characterized in that, In step S5, the graph neural network model is a hybrid model of a cascaded graph convolutional network and a deep neural network; the specific steps are as follows: S51. Normalize the node features in the graph structure dataset to eliminate dimensional differences; specifically, Z-score standardization is applied to spatial coordinates, and coordinate sequences are normalized. Calculate the standardized value using the following formula. : in, Represents the mean of the sample sequence. Indicates the standard deviation of the sample sequence; S52. Input the normalized node features and graph structure into the graph convolutional network. The graph convolutional network extracts local spatial topological features by aggregating neighbor node information through topology, and obtains the graph structure feature representation. S53. Perform an embedding average pooling operation on the graph structure feature representation to compress the features of nodes and their neighbors into dense latent vectors, which are then input into a deep neural network; the embedding average pooling is calculated according to the following formula: in, Let k represent the feature representation of a node and its neighbors after graph convolution, where k represents the number of neighboring nodes and h represents the dimension of the latent vector. This represents the dense latent vector obtained after compression; S54. The deep neural network performs regression prediction on the pooled dense latent vectors and outputs the predicted values ​​of node stress.

9. The method for full-field stress digital twin reconstruction of space docking mechanisms under sparse sensor conditions according to claim 1, characterized in that, The specific steps of step S6 are as follows: S61. Deploy the trained virtual sensor model into the digital twin system; S62. The space docking mechanism synchronizes with the digital twin model through real-time data interaction, and inputs the triaxial forces Fx, Fy, Fz and triaxial torques Mx, My, Mz collected in real time as sparse loads into the virtual sensor model; S63. The virtual sensor model reconstructs the stress of each node in the entire field based on the real-time sparse load, and visualizes the reconstructed stress of each node in the three-dimensional scene in real time, so as to realize the real-time monitoring and early warning of structural integrity during the docking process, and provide assistance for on-orbit docking safety decision-making.

10. A full-field stress digital twin reconstruction system for space docking mechanisms under sparse sensor conditions, characterized in that, include: The data acquisition module is used to acquire the fine mesh node stress of the high-precision finite element model of the space docking mechanism under various load conditions, as well as the load conditions corresponding to each load condition. The stress mapping module is used to map the stress of fine mesh nodes to coarse nodes, resulting in a coarsened node stress dataset. The dimension reduction sampling module is used to perform dimension reduction sampling on the coarsened nodal stress dataset, so that the samples are concentrated in the region of high stress intensity, and output the dimension-reduced sample set. The graph structure dataset construction module is used to construct a graph structure dataset from the dimensionality-reduced sample set. The node features in the graph structure dataset include load conditions and spatial coordinates, the node label is the stress value, and the edges between nodes are determined according to the finite element mesh connection relationship. The model training module is used to build a graph neural network model and train it using a graph structure dataset to obtain a virtual sensor model. The real-time reconstruction and monitoring module is used to deploy virtual sensor models in a digital twin system. It takes the sparse loads collected in real time by the space docking mechanism as input and outputs the predicted values ​​of stress at all nodes in the field, thereby realizing real-time reconstruction and monitoring of stress in the field.

Images